AI-Powered Recommender System for Clinical Trial Protocol Design - A Tool for Medical Practitioners

Abstract

Clinical trial endpoint selection is a complex protocol-design task that requires clinical relevance, statistical validity, regulatory awareness, and practical feasibility. In heart failure trials, this is particularly challenging because endpoint descriptions are heterogeneous, clinically nuanced, and often expressed using different terminology. This thesis investigates whether historical clinical trial data can be transformed into structured, terminology-aware representations that support secondary-endpoint recommendation for heart failure protocols. In collaboration with AstraZeneca and Evinova, the study develops an end-to end proof-of-concept pipeline. Starting from 2966 raw ClinicalTrials.gov protocol records, the dataset is reduced to 490 Phase II–III heart-failure-focused protocols containing 878 primary endpoints and 3700 secondary endpoints. Secondary end points form a reviewed hierarchy using semantic embeddings, hierarchical clustering, terminology-assisted standardization, and LLM-assisted review. Protocol and end point information is standardized against NCIt, CDISC, and LOINC for a two-stage recommendation pipeline: Stage 1 predicts relevant endpoint clusters, while Stage 2 ranks concrete secondary endpoint candidates within the predicted cluster context. The results show that hierarchical endpoint structuring provides a more interpretable and model-compatible representation than flat clustering or direct prediction over raw endpoint strings. Standardized terminology codes improved semantic consistency and contributed useful supporting features, but were most effective when combined with the reviewed hierarchy and partial endpoint context. Pairwise leave-one-out formulations were better aligned with the intended recommendation setting than direct multilabel prediction, especially for identifying missing endpoint information from a partially specified endpoint design. Full-pipeline evaluation on unseen protocols showed limited exact-match recovery, but qualitative expert review indicated that many recommendations captured clinically relevant endpoint domains, even when they were not specific enough to replace the held-out endpoint directly. Overall, the thesis demonstrates that historical clinical trial records can be reused more systematically to support endpoint-selection discussions. The proposed pipeline should be interpreted not as a production-ready clinical tool, but as a methodological foundation for AI-assisted endpoint recommendation. Future work should focus on broader therapeutic-area validation, improved terminology resources, stronger expert-labelled evaluation sets, and prospective testing with protocol designers.